diff --git a/src/external/qoa.h b/src/external/qoa.h index fc62f4765..f0f44214d 100644 --- a/src/external/qoa.h +++ b/src/external/qoa.h @@ -366,22 +366,7 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned ), bytes, &p); - for (int c = 0; c < channels; c++) { - /* If the weights have grown too large, reset them to 0. This may happen - with certain high-frequency sounds. This is a last resort and will - introduce quite a bit of noise, but should at least prevent pops/clicks */ - int weights_sum = - qoa->lms[c].weights[0] * qoa->lms[c].weights[0] + - qoa->lms[c].weights[1] * qoa->lms[c].weights[1] + - qoa->lms[c].weights[2] * qoa->lms[c].weights[2] + - qoa->lms[c].weights[3] * qoa->lms[c].weights[3]; - if (weights_sum > 0x2fffffff) { - qoa->lms[c].weights[0] = 0; - qoa->lms[c].weights[1] = 0; - qoa->lms[c].weights[2] = 0; - qoa->lms[c].weights[3] = 0; - } - + for (unsigned int c = 0; c < channels; c++) { /* Write the current LMS state */ qoa_uint64_t weights = 0; qoa_uint64_t history = 0; @@ -395,9 +380,9 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned /* We encode all samples with the channels interleaved on a slice level. E.g. for stereo: (ch-0, slice 0), (ch 1, slice 0), (ch 0, slice 1), ...*/ - for (int sample_index = 0; sample_index < frame_len; sample_index += QOA_SLICE_LEN) { + for (unsigned int sample_index = 0; sample_index < frame_len; sample_index += QOA_SLICE_LEN) { - for (int c = 0; c < channels; c++) { + for (unsigned int c = 0; c < channels; c++) { int slice_len = qoa_clamp(QOA_SLICE_LEN, 0, frame_len - sample_index); int slice_start = sample_index * channels + c; int slice_end = (sample_index + slice_len) * channels + c; @@ -405,10 +390,13 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned /* Brute for search for the best scalefactor. Just go through all 16 scalefactors, encode all samples for the current slice and meassure the total squared error. */ - qoa_uint64_t best_error = -1; - qoa_uint64_t best_slice; + qoa_uint64_t best_rank = -1; + #ifdef QOA_RECORD_TOTAL_ERROR + qoa_uint64_t best_error = -1; + #endif + qoa_uint64_t best_slice = 0; qoa_lms_t best_lms; - int best_scalefactor; + int best_scalefactor = 0; for (int sfi = 0; sfi < 16; sfi++) { /* There is a strong correlation between the scalefactors of @@ -421,7 +409,10 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned state when encoding. */ qoa_lms_t lms = qoa->lms[c]; qoa_uint64_t slice = scalefactor; - qoa_uint64_t current_error = 0; + qoa_uint64_t current_rank = 0; + #ifdef QOA_RECORD_TOTAL_ERROR + qoa_uint64_t current_error = 0; + #endif for (int si = slice_start; si < slice_end; si += channels) { int sample = sample_data[si]; @@ -434,9 +425,27 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned int dequantized = qoa_dequant_tab[scalefactor][quantized]; int reconstructed = qoa_clamp_s16(predicted + dequantized); + + /* If the weights have grown too large, we introduce a penalty + here. This prevents pops/clicks in certain problem cases */ + int weights_penalty = (( + lms.weights[0] * lms.weights[0] + + lms.weights[1] * lms.weights[1] + + lms.weights[2] * lms.weights[2] + + lms.weights[3] * lms.weights[3] + ) >> 18) - 0x8ff; + if (weights_penalty < 0) { + weights_penalty = 0; + } + long long error = (sample - reconstructed); - current_error += error * error; - if (current_error > best_error) { + qoa_uint64_t error_sq = error * error; + + current_rank += error_sq + weights_penalty * weights_penalty; + #ifdef QOA_RECORD_TOTAL_ERROR + current_error += error_sq; + #endif + if (current_rank > best_rank) { break; } @@ -444,8 +453,11 @@ unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned slice = (slice << 3) | quantized; } - if (current_error < best_error) { - best_error = current_error; + if (current_rank < best_rank) { + best_rank = current_rank; + #ifdef QOA_RECORD_TOTAL_ERROR + best_error = current_error; + #endif best_slice = slice; best_lms = lms; best_scalefactor = scalefactor; @@ -490,7 +502,7 @@ void *qoa_encode(const short *sample_data, qoa_desc *qoa, unsigned int *out_len) unsigned char *bytes = QOA_MALLOC(encoded_size); - for (int c = 0; c < qoa->channels; c++) { + for (unsigned int c = 0; c < qoa->channels; c++) { /* Set the initial LMS weights to {0, 0, -1, 2}. This helps with the prediction of the first few ms of a file. */ qoa->lms[c].weights[0] = 0; @@ -513,7 +525,7 @@ void *qoa_encode(const short *sample_data, qoa_desc *qoa, unsigned int *out_len) #endif int frame_len = QOA_FRAME_LEN; - for (int sample_index = 0; sample_index < qoa->samples; sample_index += frame_len) { + for (unsigned int sample_index = 0; sample_index < qoa->samples; sample_index += frame_len) { frame_len = qoa_clamp(QOA_FRAME_LEN, 0, qoa->samples - sample_index); const short *frame_samples = sample_data + sample_index * qoa->channels; unsigned int frame_size = qoa_encode_frame(frame_samples, qoa, frame_len, bytes + p); @@ -576,14 +588,14 @@ unsigned int qoa_decode_frame(const unsigned char *bytes, unsigned int size, qoa /* Read and verify the frame header */ qoa_uint64_t frame_header = qoa_read_u64(bytes, &p); - int channels = (frame_header >> 56) & 0x0000ff; - int samplerate = (frame_header >> 32) & 0xffffff; - int samples = (frame_header >> 16) & 0x00ffff; - int frame_size = (frame_header ) & 0x00ffff; + unsigned int channels = (frame_header >> 56) & 0x0000ff; + unsigned int samplerate = (frame_header >> 32) & 0xffffff; + unsigned int samples = (frame_header >> 16) & 0x00ffff; + unsigned int frame_size = (frame_header ) & 0x00ffff; - int data_size = frame_size - 8 - QOA_LMS_LEN * 4 * channels; - int num_slices = data_size / 8; - int max_total_samples = num_slices * QOA_SLICE_LEN; + unsigned int data_size = frame_size - 8 - QOA_LMS_LEN * 4 * channels; + unsigned int num_slices = data_size / 8; + unsigned int max_total_samples = num_slices * QOA_SLICE_LEN; if ( channels != qoa->channels || @@ -596,7 +608,7 @@ unsigned int qoa_decode_frame(const unsigned char *bytes, unsigned int size, qoa /* Read the LMS state: 4 x 2 bytes history, 4 x 2 bytes weights per channel */ - for (int c = 0; c < channels; c++) { + for (unsigned int c = 0; c < channels; c++) { qoa_uint64_t history = qoa_read_u64(bytes, &p); qoa_uint64_t weights = qoa_read_u64(bytes, &p); for (int i = 0; i < QOA_LMS_LEN; i++) { @@ -609,17 +621,19 @@ unsigned int qoa_decode_frame(const unsigned char *bytes, unsigned int size, qoa /* Decode all slices for all channels in this frame */ - for (int sample_index = 0; sample_index < samples; sample_index += QOA_SLICE_LEN) { - for (int c = 0; c < channels; c++) { + for (unsigned int sample_index = 0; sample_index < samples; sample_index += QOA_SLICE_LEN) { + for (unsigned int c = 0; c < channels; c++) { qoa_uint64_t slice = qoa_read_u64(bytes, &p); int scalefactor = (slice >> 60) & 0xf; + slice <<= 4; + int slice_start = sample_index * channels + c; int slice_end = qoa_clamp(sample_index + QOA_SLICE_LEN, 0, samples) * channels + c; for (int si = slice_start; si < slice_end; si += channels) { int predicted = qoa_lms_predict(&qoa->lms[c]); - int quantized = (slice >> 57) & 0x7; + int quantized = (slice >> 61) & 0x7; int dequantized = qoa_dequant_tab[scalefactor][quantized]; int reconstructed = qoa_clamp_s16(predicted + dequantized);